Regarding Position Encoding

[isCopyman]

May I ask if positional encoding of features is important in these Transformer based architectures? It seems that the vast majority of papers do not mention positional encoding

[6sy666]

For tabular data, the features typically do not have a fixed order, meaning that swapping the order of the raw data features does not fundamentally change the dataset. Because of this, positional encoding is generally not necessary in Transformer-based architectures when working with tabular data. The absence of positional encoding in most papers likely reflects this characteristic of tabular data.